Smart KYC with Gemini Enterprise: mid-size bank case (from 4h to 12 min)

KYC (Know Your Customer) is the regulatory process Brazilian banks use to validate identity, financial capacity and risk profile. Required by BCB Resolution 119/2021 and Circular 3,978/2020. When done manually, it is expensive: document reading, base cross-referencing, adverse media analysis, risk classification.

This post details how we built the agent that automated this process at a mid-size Brazilian bank. Client name is under NDA — the numbers are real.

Before: the starting point

The problem wasn't lack of people — it was process. Each analysis involved 7 different systems (Federal Revenue, Serasa, Sintegra, Detecta, adverse media, articles of incorporation, internal system). Chronic backlog of ~800 pending analyses.

Agent architecture

The agent was built in Gemini Enterprise Plus (US$ 39/user/month, with custom tools and Vertex AI Agent Builder access).

1. Multimodal ingest: articles of incorporation PDF, partner IDs, address proof, revenue. Gemini 2.5 Pro extracts structured data (CNPJ, partners, share capital, activity, address).
2. Query tools:
   • Federal Revenue (active CNPJ, registration status, partner structure).
   • Serasa (corporate score and individual partners).
   • Sintegra (state status).
   • PEP and sanctions list (internal + Central Bank).
   • Adverse media (structured search with temporal filters).
3. Reasoning: Gemini 2.5 Pro orchestrates tools, builds the dossier, identifies inconsistencies (share capital vs revenue, partner with restriction, activity vs history).
4. Risk classification: agent suggests low/medium/high with per-criterion justification citing evidence.
5. Output: structured PDF dossier + entry in the internal system + alerts for the human analyst.

Results at 90 days

The NPS result was a surprise. Analysts who feared replacement became fans: work shifted from copy-paste to critical analysis on a ready dossier.

BACEN governance: what had to be ready

• Documented human final decision in each case.
• Full audit log: who asked what, which tools were called, which documents consulted, which recommendation was made, who approved.
• Data residency in sa-east1 and training opt-out.
• Detailed DPIA with risk matrix (template in DPIA for Gemini Enterprise projects).
• Continuity plan: if the agent goes down, the manual process resumes without disruption.
• Quarterly bias assessment: monthly sample of 100 analyses comparing agent vs human decision.

What went wrong along the way

Error 1: generic OCR for scanned contracts

First pilot failed on 22% of old contracts. We migrated to Gemini 2.5 Pro multimodal directly on PDF — error rate dropped to 2%.

Error 2: unanchored prompt

Agent "hallucinated" classifications when the legal basis didn't cover the case. We rewrote it requiring: "if the rule for this case isn't in the attached criteria document, classify as 'requires human review' — don't invent". Verifiable hallucination → zero.

Error 3: adverse media without temporal window

Agent was treating 2009 news with the same relevance as 2026 news. Added a temporal filter (last 3 years default, 7 years for partners with restriction) and severity ranking.

Error 4: we underestimated change management

The team resisted in the first 2 weeks — saw it as a threat. Solved with a hands-on workshop: each analyst used the agent on real cases under supervision. The turn came when they saw that "saved" time became more interesting work (complex fraud analysis, in-depth analysis).

Project cost

Direct savings: 1,800 analyses × R$ 166 = R$ 298,800/month. Payback: ~5 weeks after go-live. Calculate your scenario with the ROI calculator.

Replicability

The pattern repeats in any process with (a) high volume, (b) clear but multiple rules, (c) consultation with several systems, (d) final decision that admits human review. We've already applied it to financial reconciliation, credit analysis and supplier onboarding.

Also read

• AML/PLD with Gemini Enterprise agent
• DPIA for Gemini Enterprise projects
• Corporate RAG with Vertex AI Search